Segmentation-guided diffusion model for retinal OCT speckle reduction with anatomical structure preservation

Optical coherence tomography (OCT) provides essential high-resolution imaging for retinal disease diagnosis, but inherent speckle noise degrades signal-to-noise ratio and obscures fine anatomical boundaries. Denoising diffusion probabilistic models (DDPM) offer improved image quality compared to traditional methods but suffer from high-frequency detail misalignment, compromising structure preservation. To address this limitation, we propose a segmentation-guided DDPM that incorporates retinal layer segmentation maps as conditioning information to enhance anatomical preservation. We fine-tuned a 6-class retinal layer segmentation model using the OCT5k dataset, from which four clinically relevant retinal layers were extracted as conditioning information for the diffusion model. Both standard and segmentation-guided DDPM models were trained on the PKU37 dataset using a U-Net-based architecture. The proposed method demonstrated statistically significant improvements over standard DDPM not only in image quality metrics: SNR (+ 0.515 dB, p < 0.001), PSNR (+ 0.516 dB, p < 0.001), and SSIM (+ 0.086, p < 0.001), but also in downstream classification accuracy for disease detection, improving from 0.416 (noisy) to 0.622 (segmentation-guided) compared to 0.546 (standard DDPM). External validation on an independent clinical dataset showed consistent image-quality improvement over noisy input, with qualitative preservation of the external limiting membrane. The segmentation-guided approach improves downstream classification performance while maintaining efficient single-frame processing, and may complement existing frame-averaging workflows where repeated acquisition is impractical.